AI And Its Successor: What Happens When The Tool Helps?
AIThis post was created with the assistance of artificial intelligence (AI).

🔍 Read the full analysis: AI And Its Successor: What Happens When The Tool Helps? on ThorstenMeyerAI.com

Before you orderOffer from Amazon

Get hardware and tech essentials delivered free with Prime

  • Fast, free delivery on millions of items
  • Prime Video, Amazon Music and more included
  • Member-only deals all year
Start your free Prime trial Free trial for eligible customers · Cancel anytime
As an affiliate, we earn on qualifying purchases.

TL;DR

METR’s November 5 assessment judged that Anthropic’s capabilities on well-specified tasks were improving at about twice their earlier rate, while emphasizing substantial uncertainty. Its account describes AI primarily assisting researchers, not independently running the research process. Anthropic redacted much of the evidence METR considered most informative, and bottlenecks such as experiment time and hardware could limit further gains.

METR, an independent evaluation organization, said in an assessment published November 5, 2026, that Anthropic was likely improving its models’ capabilities on well-specified tasks at about twice the earlier rate. The evaluator stressed that the estimate was highly uncertain and described the work mainly as AI assistance to researchers, rather than autonomous AI research.

METR said it had worked with internal information from August through October. Its estimate is not a precise measurement: the organization said it “would not be surprised” by a result ranging from 1.3 times to 3 times the earlier rate. Anthropic redacted much of the evidence METR considered most informative, limiting what people outside the evaluation can independently examine.

The assessment concerns improvement on well-specified tasks, not a general measure of all AI research or a finding that AI systems can independently build their successors. According to the account of METR’s work, AI was primarily helping researchers. That could include supporting parts of the work, while people still set goals, review results and decide which changes to use.

A separate example cited by the source comes from Google DeepMind: it reported that AlphaEvolve made a matrix-multiplication component used in Gemini 23% faster, reducing Gemini’s overall training time by 1%. That is a company-reported example of a targeted efficiency gain, not evidence that every improvement will compound at the same rate.

At a glance
reportWhen: METR assessment published November 5, 2…
The developmentMETR published an assessment of AI research acceleration at Anthropic, finding likely faster improvement on well-specified tasks while noting high uncertainty.
Crypto market snapshot
Fear & Greed Index
10/100 — Extreme Fear
Bitcoin BTC$64,494▼ 1.9%
Ethereum ETH$1,809▼ 1.9%
Tether USDT$0.9996▲ 0.0%
BNB BNB$590.11▼ 0.1%
USDC USDC$0.9997▲ 0.0%
XRP XRP$0.9486▼ 1.3%
Solana SOL$71.3▼ 1.9%
TRON TRX$0.3052▼ 0.6%
Live data · CoinGecko · alternative.me (24h change)

How Research Gains Could Compound

The finding matters because AI systems may help with the research and engineering that produce later systems. If models reliably help people write code, test ideas, diagnose failed experiments or assess results, a research team could investigate more possibilities with the same staff and resources. The possible feedback loop is that improved models then assist with further development. The METR assessment offers evidence of acceleration in one setting, but does not establish that this loop is already autonomous or will continue at a fixed pace.

Efficiency gains can have practical effects even when they are modest. The source offers a simplified illustration: if three independent improvements each cut a task’s resource use by 10%, the task would use about 73% of its original compute. Under the stated assumptions, the same compute budget could then support roughly 37% more such jobs. This is an illustration, not a measured result from Anthropic or Google DeepMind, and actual gains depend on whether improvements are independent and carry over to larger workloads.

For businesses and workers, the pace of AI development can shape decisions about investment, staffing and adopting new tools. The source argues that access to systems, ownership of infrastructure and how gains are shared will also depend on deployment choices and institutions, not technical capability alone. The assessment does not resolve those questions, but it adds urgency to planning for uneven and potentially changing rates of progress.

Products Worth Considering

LG UltraGear 34G600A-B 34-Inch Curved WQHD Gaming Monitor, 160Hz

LG UltraGear 34G600A-B 34-Inch Curved WQHD Gaming Monitor, 160Hz

As an affiliate, we earn on qualifying purchases.

From Research Assistance to Successors

AI research involves more than generating a promising idea. A team may need to review relevant work, inspect code, prepare data, run an experiment, compare results fairly and repeat tests before deciding whether a change is useful. A model that helps with one or several steps can raise a team’s output without taking over the whole process.

Reliability is a central constraint. An apparent improvement might result from an easier test, unequal computing resources or a result that disappears with a different random seed. Researchers need to identify such problems, check whether evidence supports a conclusion and decide whether a change should be adopted. The account says that full automation would involve systems choosing what to investigate as well as proposing, implementing and evaluating experiments; the METR assessment described primarily assistance, not that full process.

Google DeepMind’s AlphaEvolve example illustrates how a narrow engineering change can affect a larger workload: the company said a matrix-multiplication component used in Gemini became 23% faster, while total training time fell 1%. The difference between those figures shows why improvements to one component should not be treated as equivalent to the same percentage reduction in an entire system’s costs.

“would not be surprised”

— METR

Evidence and Bottlenecks Still Matter

The public account leaves important limits. Anthropic redacted much of the evidence METR found most informative, so outside readers cannot fully assess the underlying material from the information provided. METR’s estimate also carries a wide range, and the finding applies to capabilities on well-specified tasks; it should not be read as a firm estimate of the pace of all AI development.

It is also unclear how much of the observed acceleration would persist as tasks become less defined or systems grow larger. The source identifies possible brakes: experiment time and hardware, the gap between small tests and large systems, the choice of research direction, continuity of work and data quality. It does not quantify how much any one of these factors currently slows progress.

The source’s broader possibility of rapid progress toward superintelligence is an interpretation, not a conclusion established by the METR assessment. The assessment indicates likely acceleration in a specific research setting, with substantial uncertainty; it does not establish a timetable for superintelligence, the degree of future automation or the economic effects.

What Future Assessments Must Show

The next evidence readers need is whether AI assistance continues to improve research output and whether those gains hold up under independent checks. Further assessments could clarify how much work systems perform, how much supervision researchers provide, and whether gains seen on well-specified tasks transfer to larger, more complex experiments.

For now, the source recommends planning for a pace of improvement that can change rather than assuming either steady acceleration or a fixed ceiling. Decisions about access, infrastructure ownership and distribution of benefits will depend on how systems are deployed and governed. The available material does not specify a next METR assessment date or a further Anthropic milestone.

Key Questions

What did METR conclude about Anthropic?

METR judged that Anthropic’s capabilities on well-specified tasks were likely improving at about twice the earlier rate. It emphasized substantial uncertainty, saying it would not be surprised by a rate between 1.3 and 3 times the earlier one.

Did METR find that AI is independently conducting AI research?

No. The account says the work METR described was primarily AI assistance to researchers. It does not establish that AI systems were independently choosing research goals and carrying out the full process.

Why can a small efficiency gain matter?

A reduction in the resources needed for one task could let a team run more tasks within the same budget. The source’s example of three independent 10% savings is a simplified illustration, not a reported outcome from the METR assessment.

What makes the finding difficult to verify?

Anthropic redacted much of the evidence METR considered most informative. METR also described its estimate as highly uncertain, and the public account does not provide enough detail for outsiders to independently assess all of the underlying evidence.

Does this report establish when superintelligence will arrive?

No. The assessment concerns the pace of improvement on well-specified tasks in Anthropic’s research setting. It does not provide a timetable for superintelligence or prove that current acceleration will continue.

Source: ThorstenMeyerAI.com

Nothing in this article is financial or investment advice. Cryptocurrency and precious-metal investments carry significant risk — do your own research and consider a licensed advisor.
You May Also Like

10 AI Trends That Will Define The Next Decade

An analysis of the top 10 AI trends expected to define the next decade, based on expert insights and industry developments.

Why Batteries Are Joining The 800-Volt DC Shift In AI Data Centers

AI data center plans are bringing 800-volt DC and batteries into focus as operators confront rising rack demand and slow grid connections.

Dubai’s Crypto Tower: A 17-Story Landmark for Blockchain Innovation

Journey into Dubai’s Crypto Tower, where groundbreaking blockchain innovations await—discover how this iconic structure is transforming the future of cryptocurrency.

Secure Non‑Custodial Limit Order Protocols

Many secure non-custodial limit order protocols revolutionize trading by enhancing safety; discover how they safeguard your assets and redefine flexibility.